Sub-agent models and effort

RPI skills run in the main session. Some skills launch a sub-agent for a focused task. The session's coding agent controls the sub-agent model and effort rules.

The two main rules are:

• Claude Code: Most RPI sub-agents use a fixed Claude model family. Only agents marked inherit use the parent model.
• Codex: Every sub-agent uses the parent GPT model. Most use low effort. The outline implementer uses the parent effort.

Model and effort are separate settings. A sub-agent can use the parent model at lower effort. It can also use another model family with the session's Claude thinking settings.

For the role of sub-agents in each phase, read How workflow phases fit together. For each skill's exact behavior, read the skills and workflows reference.

Claude Code

Model selection

HumanLayer registers these RPI sub-agents for Claude Code sessions:

Sub-agent	HumanLayer UI	Claude Code Terminal UI
codebase-analyzer	Sonnet, or Haiku with Faster Research Subagents	Sonnet
codebase-locator	Sonnet, or Haiku with Faster Research Subagents	Sonnet
codebase-pattern-finder	Sonnet, or Haiku with Faster Research Subagents	Sonnet
web-search-researcher	Sonnet	Sonnet
implementer-agent	Sonnet	Sonnet
implementation-reviewer	Sonnet	Sonnet
outline-implementer-agent	Parent model (inherit)	Parent model (inherit)

sonnet, haiku, and inherit are Claude Code model aliases. Claude Code resolves Sonnet and Haiku to the model behind each alias. inherit uses the exact main-session model.

No RPI sub-agent uses a fixed Opus model. If the parent uses Opus, the skill runs on Opus. Only outline-implementer-agent also uses Opus. The other named sub-agents use Sonnet or Haiku.

Fable follows the same rule. With a Fable parent, outline-implementer-agent uses Fable. The fixed-model sub-agents use Sonnet or Haiku.

A Haiku parent can therefore launch a Sonnet implementer. A fixed child model does not use the parent model family.

Effort

For HumanLayer UI sessions, HumanLayer sends the selected effort to Claude Code. It does not set a separate effort for each RPI sub-agent.

Claude Code manages child thinking separately from the model alias. The current bundled version sends the main session's extended-thinking settings to the sub-agent. Terminal UI sessions use the effort in the Claude CLI configuration.

When no effort is set, Claude Code uses the model default. A direct SDK learning test used an Opus 5 outline-implementer-agent with model: inherit. Each child request used medium effort when the test set no effort.

The exact rules are:

• model: inherit controls the model. It does not control effort.
• The main Claude session supplies the child's thinking settings.
• HumanLayer does not add an RPI-specific Claude effort setting.
• Claude Code and the selected model set the final child effort when the session does not set one.

Faster Research Subagents

The Faster Research Subagents switch is in Settings -> Experiments. The user's setting and the organization's setting must both allow it.

When enabled, it changes only these three agents from Sonnet to Haiku:

• codebase-analyzer
• codebase-locator
• codebase-pattern-finder

It does not change:

• the parent session model
• web-search-researcher
• either implementation agent
• implementation-reviewer
• Codex sub-agents
• sub-agents in Claude Code Terminal UI

Use the switch when code search speed and cost are more important than Sonnet's added analysis. Turn it off when the research task needs Sonnet's stronger code analysis.

Codex

Codex sessions run through CodeLayer. Every child uses the exact GPT model selected for the parent. Changing the parent GPT model changes all its sub-agents.

Sub-agent	Model	Effort
general-purpose	Same model as parent	Low
bash	Same model as parent	Low
codebase-analyzer	Same model as parent	Low
codebase-locator	Same model as parent	Low
codebase-pattern-finder	Same model as parent	Low
web-search-researcher	Same model as parent	Low
implementer-agent	Same model as parent	Low
outline-implementer-agent	Same model as parent	Same effort as parent

For example, a GPT parent at high effort starts research and standard implementation sub-agents at low. The children use the same GPT model. The outline implementer uses that model at high.

If the parent uses low, the standard child stays at low. HumanLayer does not increase the effort.

Codex Fast Mode is separate from Faster Research Subagents. Fast Mode uses the Codex session setting. Faster Research Subagents changes only the three Claude Code research agents listed earlier.

Parent examples

Parent session	Research sub-agent	Standard implementer	Outline implementer
Claude Code, Opus	Sonnet or Haiku	Sonnet	Opus
Claude Code, Fable	Sonnet or Haiku	Sonnet	Fable
Claude Code, Sonnet	Sonnet or Haiku	Sonnet	Parent Sonnet model
Claude Code, Haiku	Sonnet or Haiku	Sonnet	Parent Haiku model
Codex, GPT at high effort	Same GPT at low	Same GPT at low	Same GPT at high
Codex, GPT at low effort	Same GPT at low	Same GPT at low	Same GPT at low

What this page does not cover

The HumanLayer Library Researcher uses a separate managed cloud workflow. It does not use the local parent model. Faster Research Subagents does not affect it.

Project-defined and user-defined sub-agents can set their own model and effort. Check each agent definition instead of using the defaults on this page.

For Claude's model, inherit, and effort fields, see the Claude Agent SDK sub-agent reference.